Biomarker development is a high priority area of research for NIH. However, standards for statistical design and evaluation of biomarker studies are not well developed, especially when compared with other medical fields such as therapeutics and epidemiology. This grant proposes to develop statistical methods for several important problems in biomarker research focusing on use of biomarkers for screening, diagnosis, prognosis and risk prediction. Aim 1 concerns improvement in prediction (or detection) of disease when a novel biomarker is added to an existing set of clinical predictors/markers. We will investigate: (i) The statistical characteristics of a new marker relative to existing predictors that lead to large versus small improvements in performance. Practical implications will be developed for selecting marker panels for validation from large sets of candidate biomarkers in discovery research; (ii) Techniques for making statistical inference about measures that quantify improvement in performance, improving upon existing methods that are flawed; and (iii) Designs for studies to estimate improvement in performance with a focus on the practice of choosing controls to match cases in regards to clinical predictors or existing markers. We will develop methods for estimation with a matched design and investigate if the design leads to more efficient use of data. The goal of aim 2 is to develop a framework for simultaneously estimating relative risks associated with a biomarker and its performance for classification. These two tasks are currently done separately leading to disjointed, and sometimes inconsistent, results. In addition to providing a more coherent approach to analysis, we believe that wide availability and familiarity of methods for relative risk regression will enable researchers to do more sophisticated analyses of biomarker performance than is currently done in practice. In particular, we propose to develop methods in this framework for comparing biomarkers, for evaluating factors that affect biomarker performance and for meta-analysis that combines data from different studies or subpopulations. Methods will be evaluated using cancer biomarker datasets from collaborations with investigators in the Early Detection Research Network, as well as collaborations on HIV research, emergency medicine research and fertility research. Mathematical theory and simulation models will be used to evaluate and compare statistical techniques. Software will be written in Stata and R statistical packages and will be made available via our Diagnostics and Biomarker Statistical Center website, http://labs.fhcrc.org/pepe/dabs/.